Shipping real software with AI agents: the harness beats the model
After a year of shipping with agents, the lesson is boring and important. The harness around the model, the tools, the context, the verification loop, matters more than which model you picked.
Agentic coding moved from smarter autocomplete to a closed loop: the agent reads a task, plans, edits several files, runs the tests and the build, reads the failures, and iterates. Average session length grew from a few minutes in early 2025 to over twenty in 2026, which is really a proxy for how much work an agent can carry before it needs a human.
The harness beats the model
The most useful finding from teams running agents at scale is deflating: most failures are not capability problems, they are configuration problems. A weaker model with a good harness, meaning the right tools, careful context assembly, a verification loop and guardrails, will beat a stronger model with none. If your agent keeps failing, fix the harness before you swap the model.
Spec-driven development
The discipline that makes agents shippable is to externalize intent as artifacts. The loop that stuck is Specify, Plan, Tasks, Implement, each with a human checkpoint. A project constitution file carries the standing context so you are not re-explaining the stack every session.
# AGENTS.md ## Stack Next.js 16, TypeScript, Payload 3, MongoDB. ## Rules - Server Components by default; "use client" only at the leaves. - Run pnpm lint and pnpm typecheck before every commit. - No secrets in code. Read config from the environment.
The failure modes are real
Agents reliably find the right file but often miss the exact lines to change. End-to-end tests break in a handful of predictable ways: brittle selectors, implicit timing assumptions, schema drift. And a meaningful slice of generated code, studies put it near a quarter to a third of snippets, carries a security weakness. Independent review, not the agent reviewing its own PR, is non-negotiable, because self-review validates against the same model that wrote the bug.
The connective tissue is MCP, the Model Context Protocol, which became the integration standard the way USB-C did for cables. It is not just a demo: one large engineering org reported tens of thousands of monthly tool invocations through its internal MCP setup and thousands of engineering hours saved.
The through-line is that engineering judgment moved upstream. Requirements, architecture and review matter more than typing. The people getting the most out of agents are the ones who write the clearest specifications, then let the machine do the mechanical part.